Accelerating Scientific Discovery with Autonomous Goal-evolving Agents

1 Introduction

• SAGA’s outer loop proposes and implements objectives, optimizes candidates, and analyzes results, while its inner loop searches under current objectives; scientists can collaborate in co-pilot, semi-pilot, or autopilot modes.The framework addresses the difficulty of specifying objectives upfront: fixed objectives can be exploited by optimization, and objective choices and weights are themselves combinatorial.Source: §1 · p. 3

2.1 SAGA for Antibiotic Design

• Four compounds inhibited E. coli growth by >80% at 128 µg/mL with PMB; compound 8 showed minimal cytotoxicity at its MIC and a Tanimoto similarity of 0.28 to known antibiotics.Across three automation modes, SAGA balanced biological objectives with chemical filters, while dynamic objective evolution addressed chemical-quality failures and improved candidate passing rates.Source: §2.1 · pp. 4, 6

2.2 SAGA for Nanobody Design

• Three SAGA-designed nanobodies bound PD-L1, with a lowest KD of 300 nM; its composite score separated binders from non-binders (p = 0.03), using 2–5 times fewer oracle calls than baselines.The binders’ CDR3 sequences shared less than 20% similarity with comparison nanobodies and antibodies, and nanobody A2 was predicted to have a novel alpha-helical CDR3.Source: Figure 3 · p. 7, §2.2 · p. 8

2.3 SAGA for Functional DNA Sequence Design

• SAGA surpassed DNA-design baselines by 19.2–176.2% across metrics; under controlled objectives, it improved MPRA specificity by at least 48.0%, motif enrichment by 47.9%, and sequence stability by 1.7%.Human feedback helped SAGA add motif-enrichment and GC-content objectives; autopilot achieved comparable HepG2 specificity and improved sequence diversity and stability. Designed enhancers also showed HepG2-specific CAGE-seq and DNase-seq signals in independent predictions.Source: §2.3 · pp. 9, 11

2.4 SAGA for Inorganic Materials Design

• SAGA achieved a more balanced overall score distribution for permanent magnets than baselines, while TextGrad and SAGA-Opt failed to balance key properties such as stability and Curie temperature.TextGrad struggled to optimize magnetocrystalline anisotropy energy without rare earth elements; SAGA-Opt over-optimized saturation magnetization and HHI without reasonable thermodynamic stability or high Curie temperature.Source: §2.4 · p. 13
• For superhard materials, all three SAGA modes achieve high scores across metrics, while fixed-objective TextGrad performs much lower on hardness and elastic modulus.SAGA begins by maximizing bulk and shear moduli, then refines objectives; more than 90% of proposed crystals contain light elements, and over 75% are transition-metal carbides, nitrides, or borides.Source: §2.4 · p. 13
• SAGA proposes material objectives aligned with task requirements, including thresholds, rare-earth penalties, and composite targets balancing competing magnetic properties.For permanent magnets, it introduces HHI below 1500 and Ehull below 0.1 eV/atom thresholds; semi-pilot adds the Elemental Criticality Index, while autopilot adds K1/Ms. In superhard-material design, modes prioritize Vickers hardness, elastic modulus, and Pugh ratio, with autopilot comparable across five metrics.Source: §2.4 · pp. 13–14

2.5 SAGA for Chemical Process Design

• SAGA improves chemical separation designs over purity-only RL by raising capital-cost and material-flow-intensity scores while maintaining high, near-ideal purity.The comparison covers butanol/water separation across feed compositions; gains in costs and flow intensity can accompany slight purity variation because objectives may conflict and are sensitive to their weights.Source: §2.5 · pp. 14, 16, Figure 6 · p. 15
• SAGA’s chemical-process agent proposes objectives such as complexity, component recovery, and material efficiency, and uses text descriptions of flowsheets to analyze design potential.The descriptions organize feed streams, unit operations, process flow, and product streams; SAGA implements objectives and scoring functions across co-, semi-, and autopilot levels.Source: §2.5 · pp. 14, 16

3 Discussion

• Across five tasks, SAGA’s evolving objectives support experimentally validated antibiotic and nanobody discoveries, novel functional DNA enhancers, high-quality superhard materials, and more relevant chemical flowsheets.The antibiotic hit showed activity against E. coli without cytotoxicity in human cell lines; nanobody testing confirmed three binders among 24 candidates (KD as low as 300nM), and the composite objective distinguished binders from non-binders (p = 0.03).Source: §3 · p. 16
• SAGA supports interactive scientific workflows by letting people revise objectives and intervene, while allowing autonomous operation when objectives and evaluation pipelines are mature.Intermediate analysis can expose synthesis liabilities in antibiotic candidates or off-target rates and missing cell-type-specific motifs in DNA designs, informing subsequent objectives.Source: §3 · pp. 16–17
• SAGA currently depends on computationally verifiable objectives and high-level goals that often predefine the design space; broader autonomy requires human or lab feedback and automatic design-space formulation.The discussion frames its inner optimization loop as “thinking fast” and its outer objective-evolution loop as “thinking slow,” with physical experiments and broader problem formulation as next steps.Source: §3 · p. 17

4.1.1 Overview

• SAGA alternates objective evolution in an outer loop with candidate optimization in an inner loop, guided by a planner, implementer, optimizer, and analyzer.Users provide a natural-language goal and may also supply context, initial objectives, or candidates; the system returns hypotheses intended to satisfy the goal.Source: §4.1.1 · pp. 17–18, S1.1 Framework Design · p. 35

4.1.2 Core Modules

• The planner proposes measurable objectives, the implementer creates executable scoring functions, the optimizer searches candidates, and the analyzer reports progress and can stop when success criteria are met.The default optimizer is an LLM-based evolutionary algorithm; SAGA also supports other algorithms, while unreliable or infeasible objectives can be sent back for revision.Source: §4.1.2 · p. 18, S1.1.1 Key Concepts · p. 35, S1.1.2 Modules · p. 36

4.1.3 Autonomy Levels

• SAGA offers co-pilot, semi-pilot, and autopilot modes, ranging from close scientist collaboration to autonomous operation of all four modules.In co-pilot mode scientists review planner and analyzer proposals; in semi-pilot mode they guide planning and strategic decisions; autopilot supports exploration when constraints are specified and system trust is established.Source: §4.1.3 · pp. 18–19

4.2 Task Configurations

• Antibiotic design optimizes activity, novelty, synthesizability, toxicity, and known-motif avoidance, while nanobody design targets PD-L1 affinity, interface quality, and sequence developability.Antibiotic candidates use an LLM evolutionary search with product-aggregated scores and cluster-based selection; nanobodies use a caplacizumab-derived scaffold, genetic crossover and mutation, Boltz2 evaluation, and CDR-diversity filtering.Source: §4.2 · pp. 19–20
• DNA design targets cell-type-specific enhancer and promoter activity, while materials and chemical-process tasks use domain-specific objectives and optimizers.DNA candidates are selected for target-cell expression and diversity; materials tasks cover low-risk permanent magnets and superhard materials, while process design uses an RL agent to build separation flowsheets and iteratively refine objectives.Source: §4.2 · pp. 20–21

4.3 Task Evaluations

• Evaluation uses task-specific computational criteria, including biological activity and drug-likeness for antibiotics, binding and developability for nanobodies, expression selectivity for DNA, and physical or process properties for materials and flowsheets.Evaluations are unseen during SAGA’s online run; nanobody metrics are assessed with AlphaFold3 and Boltz2, and materials properties with DFT calculations.Source: §4.3 · pp. 21–22

4.4 Experimental Validation

• Antibiotic validation measures E. coli growth inhibition and human-cell cytotoxicity, while nanobody validation measures PD-L1 binding kinetics and equilibrium affinity by BLI.Antibiotic dose-response curves and MICs are determined for compounds 4, 8, 9, and 27; the nanobody assay uses concentration-series sensorgrams and kinetic and steady-state analysis.Source: §4.4 · pp. 22–23

Code Availability

• The SAGA code is available under the MIT license.The framework structures information around candidates, populations, objectives, and scoring functions, with scoring-function execution isolated in Docker containers.Source: Code Availability · p. 23, S1.1.1 Key Concepts · p. 35

S1.1.2 Modules

• The Analyzer combines score trends with tool-assisted candidate investigations to report progress, issues, and recommendations, and decides whether optimization should continue or stop.A selector can also assess and rank candidates from the full run against the discovery goal, rather than relying only on final-iteration scores.Source: S1.1.2 Modules · p. 37, S1.2 Execution Workflow · p. 38

S1.2 Execution Workflow

• SAGA cycles through planning, scoring-function implementation, optimization, and analysis, then saves results and can retrospectively select candidates across all iterations.Failed objective implementations trigger Planner revisions, while optimization can inject random candidates and uses an evolutionary algorithm by default; the Analyzer assesses termination and recommends whether to continue.Source: S1.2 Execution Workflow · pp. 37–38

S1.3 Experimental Setups

• Experiments use task-specific optimizer backbones, GPT-5 for planning and objective matching, and Claude Code with Claude Sonnet 4.5 for new scoring functions, candidate investigation, and selection.The Analyzer uses GPT-5 for comprehensive reports and termination decisions; each task starts with a high-level goal, context, initial objectives and scoring functions, and randomly initialized candidates.Source: S1.3 Experimental Setups · p. 38

S1.6 Comparing SAGA with Other AI Scientists

• Supplementary Table S1.1 compares representative AI scientific discovery systems by agent type, tool use, human involvement, research autonomy, self-evolution, discovery target, and domain generality.The text points readers to the table for comparisons between SAGA and other AI scientists.Source: Table S1.1 · p. 39, S1.6 Comparing SAGA with Other AI… · p. 39

S1.4 Extensibility and Customization

• Users can extend SAGA by replacing modules, adding task-specific scoring functions, and configuring module choices and loop parameters without changing code.For example, the default optimizer can be replaced with a genetic algorithm or Bayesian optimization, while user-provided scoring functions can be reused by the Implementer.Source: S1.4 Extensibility and Customization · p. 39

S1.5 SAGA Reduces to the Optimizer agent (SAGA-Opt)

• With its outer loop disabled, SAGA reduces to SAGA-Opt, which uses fixed objectives and the default evolutionary optimizer, similar to AlphaEvolve.SAGA-Opt experiments use the default version without specialized design.Source: S1.5 SAGA Reduces to the Optimizer… · p. 39

S1.7 Adapting SAGA to a New Problem

• Adapting SAGA requires specifying a discovery goal, choosing an optimizer, optionally supplying initial objectives and scoring functions, and configuring and running the outer loop.Goals can include a separate context block; human-provided scoring functions are reused and can guide implementation of new ones. In Co-pilot and Semi-pilot modes, scientists can review or edit Planner proposals or Analyzer reports, and the run record preserves candidates, scores, objectives, reports, and implementations.Source: S1.7 Adapting SAGA to a New Problem · p. 40

S2.1 Supplementary Figures

• The supplementary figures cover external-evaluation pass rates, property distributions, optimization objectives across epochs, and changes in pass rates across iterations.Source: Figure S2.1 · p. 41, Figure S2.2 · p. 42, Figure S2.3 · p. 43

S2.2.1 Objectives, metrics and baselines

• Evaluation defines pass rate as the share of molecules meeting selected thresholds, which are also used to select candidates for future experiments.The benchmark includes four representative prior approaches spanning LLM-driven optimization, science language models, and reinforcement-learning molecular generators.Source: S2.2.1 Objectives, metrics and baselines · p. 47

S2.2.2 High-level Goal

• The antibiotic-design goal combines predicted antibacterial activity with low toxicity, structural novelty, avoidance of problematic motifs, and drug-likeness properties.Source: S2.2.2 High-level Goal · p. 48, S2.2.3 Context Information · p. 48

S2.2.3 Context Information

• The task instructions require candidate-wise objective scores to be normalized to [0, 1] with higher values better, and filter scores to return 1 for pass and 0 for fail.Source: S2.2.3 Context Information · p. 48

S2.3.1 Analyses of optimization convergence in the experiment

• Objectives proposed by the three SAGA modes converge toward similar types early, and objectives added after iteration 1 do not obviously improve molecule quality.The authors attribute similar mode performance to this convergence and suggest early objectives and the Analyzer’s comprehensive reports help limit further agent runs.Source: S2.3.1 Analyses of optimization… · pp. 48–49

S2.3.2 Ablation studies

• The optimizer improves its corresponding objective across iterations, while the Implementer produces five scorers with low MSE and high correlation to human-implemented functions.The Implementer comparison used 10,000 randomly sampled Enamine REAL molecules.Source: S2.3.2 Ablation studies · p. 49

S2.3.3 SAGA-Autopilot new proposed objectives after iteration 1.

• SAGA-Autopilot proposed a custom QED score that penalizes excess rotatable bonds, low Csp3 fraction, and molecular weight above 400.The proposed score was intended to address iteration 1’s QED decline and negative activity correlation.Source: S2.3.3 SAGA-Autopilot new proposed… · p. 49
• A metabolic stability score penalizes primary aliphatic amines, morpholine rings, unprotected phenols, and aliphatic aldehydes or ketones.This objective followed observations that many high-activity iteration 1 molecules contained primary amines or morpholine rings.Source: S2.3.3 SAGA-Autopilot new proposed… · pp. 49–50

S3.1 Supplementary Figures

• The cross-predictor comparison evaluates AF3 and Boltz2 predictions for 15 SAGA designs, 15 BoltzGen nanobodies, and 15 Germinal designs across binding, stability, interface, sequence, epitope-contact, and developability metrics.Scatter plots indicate predictor parity and report Pearson correlations; box plots summarize liability and hydrophobicity distributions.Source: Figure S3.1 · p. 50
• Three SAGA-designed PD-L1 nanobodies bind experimentally with dissociation constants from 300 to 400 nM.The two additional binders shown in the supplementary figure have measured KD values of 390 nM and 395 nM.Source: Figure S3.2 · p. 51

S3.2.1 Objectives, metrics, and baselines

• Evaluation uses AlphaFold3 and Boltz2 structure predictions and compares SAGA with BoltzGen and Germinal, two distinct nanobody-design paradigms.The methods are also compared across eight normalized evaluation dimensions and computational efficiency.Source: S3.2.1 Objectives, metrics, and… · pp. 52–54

S3.2.2 High-level Goal

• The nanobody objectives target PD-L1 binding confidence, structural stability, interface quality, CDR–epitope contacts, sequence–structure compatibility, and developability.The stated goal is to design high-affinity PD-L1 nanobodies for therapeutic applications in cancer immunotherapy.Source: S3.2.3 Context Information · p. 55

S4.1 Supplementary Figures

• Supplementary materials include analyses across design modes, cross-modality predictions, baseline comparisons, ablations, and promoter comparisons across five cell lines.Source: Figure S4.1 · p. 56, Figure S4.2 · p. 57, Figure S4.3 · p. 58

S4.2.1 Objectives, metrics, and baselines

• The evaluation combines held-out expression specificity metrics with motif enrichment, sequence diversity, and GC-content stability, and benchmarks against regLM, TextGrad, and Biomni.The HepG2 design objectives initially maximize HepG2 MPRA prediction and minimize K562 and SKNSH MPRA predictions.Source: S4.2.1 Objectives, metrics, and… · pp. 62–63

S4.2.2 High-level Goal

• The HepG2 enhancer task seeks 200-base-pair sequences with high HepG2 expression and minimized expression in other cell lines.Source: S4.2.2 High-level Goal · p. 64, S4.2.3 Context Information · p. 64

S4.3.1 Experiments for all cell lines

• SAGA-generated HepG2 enhancers show obvious cell-type specificity in MPRA expression.Source: S4.3.1 Experiments for all cell lines · p. 64
• SAGA outperforms all baselines by at least 16.3% for K562-specific enhancers and at least 1.5% for SKNSH-specific enhancers.The reported comparisons preserve MPRA-based specificity and biology-driven metrics.Source: Figure S4.5 · p. 60, S4.3.1 Experiments for all cell lines · p. 64
• SAGA also produces promoters described as having good quality and specificity across K562, HepG2, GM12878, SKNSH, and A549.The comparisons are summarized in Supplementary Figure S4.7.Source: S4.3.1 Experiments for all cell lines · pp. 64–65

S4.3.2 Ablation Studies

• The Implementer’s HepG2-versus-other-cell-line objective shows high, significant correlation with the human-proposed scoring function.The authors interpret this as evidence that the Implementer constructs objectives reflecting biological ground truth.Source: S4.3.2 Ablation Studies · p. 65
• The optimizer identifies HepG2-specific enhancer patterns using MPRA differences alone, but cannot balance competing biological constraints without additional modules.The authors motivate integrating Planner, Analyzer, and Implementer for more complex objectives and final candidates.Source: S4.3.2 Ablation Studies · p. 65

S5.1 Supplementary Figures

• Figure S5.3 compares Co-pilot property distributions and qualifying-structure counts with MatterGen, using magnetic density above 0.2 Å−3 and HHI below 1500 as targets.The count comparison covers 200 DFT property calculations; Figure S5.2 depicts the two-property Co-pilot target.Source: Figure S5.2 · p. 66, Figure S5.3 · p. 67

S5.2.1 LLM-based evolutionary algorithm for materials design

• The materials optimizer uses LLM-generated formulas, DiffCSP crystal structures, MatterSim geometric optimization, objective scoring, and Pareto-front selection from generated and parent crystals.The initial crystal formulas are randomly sampled from Materials Project; LLM crossover operations generate new formulas from parent formulas and objective scores.Source: S5.2.1 LLM-based evolutionary… · p. 68

S5.2.2 Objectives, metrics, and baselines

• Permanent-magnet design targets magnetic density above 0.2 Å−3 and HHI below 1500, while superhard-material evaluation includes mechanical properties and energy above hull.For magnets, SAGA adds HHI minimization after first maximizing magnetic density; 200 structures from each iteration are randomly selected for DFT verification.Source: S5.2.2 Objectives, metrics, and… · pp. 68–69
• MatterGen and TextGrad serve as inorganic-material baselines; MatterGen offers unconditional and property-conditioned generation, while TextGrad optimizes formulas from the same objectives.The experiments use three replicates; the materials-design optimization starts from 1,000 random formulas and runs for up to ten iterations with early stopping.Source: S5.2.2 Objectives, metrics, and… · p. 69

S5.2.3 High-level Goal

• The high-level goals are to generate permanent-magnet crystals with high magnetic density and low supply-chain risk, and superhard materials for ultra-precision cutting tools.Source: S5.2.3 High-level Goal · p. 69

S5.2.4 Context Information

• For both tasks, the agent is instructed to propose common, well-defined material properties relevant to the design goal and not considered in previous iterations.Source: S5.2.4 Context Information · p. 70

S5.3 Material Property Evaluation

• Material properties are evaluated with DFT procedures, including magnetic density as unit-cell magnetization divided by volume and HHI as a composition-weighted measure of geological-reserve concentration.The supplement also describes energy-above-hull and elastic-modulus calculations; HHI below 1500 is classified as low supply-chain risk.Source: S5.3 Material Property Evaluation · p. 70

S6.1 Supplementary Figures

• SAGA’s process representations include a matrix used by the RL optimizer, an automatically generated text form for SAGA, and a serialized form for implementing objectives.The figure contrasts these representations with a manually created visual illustration.Source: Figure S6.1 · p. 71

S6.2.1 Objectives, metrics, and baselines

• SAGA starts chemical-process design with maximizing product purity and evaluates product purity, capital costs, and material-flow intensity as normalized objectives.Purity rewards product streams above molar composition x > 0.9, with extra rewards above 0.95, 0.97, and 0.99; the other metrics penalize unit-operation costs and impractical small or excessive recycle streams.Source: S6.2.1 Objectives, metrics, and… · p. 73
• The RL optimizer builds flowsheets through hierarchical discrete and continuous decisions over four unit operations, including material flows and recycles.It uses matrix flowsheet representations and short-cut process models; the purity-only RL agent is the baseline, run with three random seeds, and the outer loop runs for three iterations with three experiment replicates.Source: S6.2.1 Objectives, metrics, and… · pp. 73–74

S6.2.2 High-level goal

• The case study targets flowsheets separating binary azeotropic mixtures across butanol/water feed compositions, using an RL optimizer trained from scratch in each objective iteration.The evaluation spans composition x ∈ [0, 1] in steps of 0.02; the prompt cautions that training is costly and may be unstable, particularly with excessive objectives or sensitive weights.Source: S6.2.1 Objectives, metrics, and… · p. 73, S6.2.2 High-level goal · p. 74, S6.2.3 Context · p. 74

S6.3 Challenges in chemical process design

• Chemical-process design is challenging because text representations encode topology and unit specifications, simulations can be costly and parameter-sensitive, and objectives can conflict while RL may yield impractical designs.The study therefore uses a simple separation case to assess whether SAGA can identify process issues and suitable objectives, while expecting variation from RL sensitivity and instability.Source: S6.3 Challenges in chemical process… · pp. 74–75

S7.1 Antibiotic Design

• 186× improvement in mean antibacterial activity (0.0018 →0.3356) was reported in the Co-pilot example, but only 6.7% of candidates met all quality criteria.The analyzer’s quality criteria were identified as its own.Source: Overview · p. 75
• 122× improvement in antibacterial activity (mean 0.0984) was reported in the Semi-pilot example, while safety declined by 4.6%.All candidates passed both the motif and molecular-weight filters.Source: Overview · p. 77
• 100× improvement in mean antibacterial activity (0.0008 →0.1051) was reported in the Autopilot example; safety and drug-likeness also regressed.The same example reports safety 0.751 (−0.071) and QED drug-likeness 0.725 (−0.054).Source: Overview · p. 78
• A metabolic-stability prediction objective using CYP450 interaction prediction or structural-alert scoring was proposed to avoid compounds with strong in-vitro activity but poor in-vivo stability.The passage describes this as a proposed objective, not a reported experimental result.Source: Overview · p. 79

S7.2 Nanobody Design

• Mean protein_iptm was 0.809 and mean ptm was 0.855, alongside a mean liability_score of 145.8, indicating strong predicted binding confidence but sequence-level liabilities.The initial designs also showed interface spread, including mean ∆SASA of approximately 894 Å2 and min_pae of 1.93 Å.Source: Overview · pp. 79–80
• 100/100 candidates passed the hard CDR3 helix constraint, but cdr_plddt remained 74.66 ± 3.26 versus binder_plddt 87.37 ± 1.51, and mpnn_expected_recovery was ≈0.37.The analysis identified CDR3 structural uncertainty as the main bottleneck despite high, stable binder confidence.Source: Overview · p. 81

Planner (refined)

• A 65% CDR3 α-helix pass rate left 35% of candidates failing, while marginal ipTM/pTM decline indicated tension between helix enforcement and binding objectives.CDR3 pLDDT improved modestly, but the helix-rule margin averaged 0.566, with many designs near constraint boundaries.Source: Overview · pp. 83–84

Iteration 3+

• Liability score fell by −14.7 and salt bridges rose by +0.4, while H-bonds fell by −0.3, ∆SASA by −20 Å2, and protein_iptm was essentially flat (−0.001).The changes followed objective-weight rebalancing; the gains in developability and electrostatics coincided with regressions in hydrogen bonding and buried surface area.Source: Overview · p. 86
• Liability score improved (−14.7) with a tighter distribution (153 ± 3.4 vs 168 ± 10.6), while H-bonds and ∆SASA declined and salt bridges improved.These changes suggest optimization favored sequence cleanliness and electrostatics at the cost of interface packing; narrowed liability and confidence distributions raised early-convergence concerns.Source: Overview · p. 86

S7.3 Functional DNA Sequence Design

• Iteration 1 selected 282 of 5,000 DNA candidates for HepG2-specific enhancers and improved objectives and specificity margin, but off-target activity remained positive and some candidates had high off-target activity.The objectives lacked motif-enrichment and stability information, while variance contraction could signal narrowing exploration and diversity loss.Source: Overview · p. 87

Strategic Recommendations (after human modification)

• Recommendations added motif-enrichment and stability objectives, specificity-margin pressure, and penalties or thresholds to curb motif regression and unstable outliers.The proposed updates also preserve the base objectives and off-target criteria while clarifying HepG2 versus off-target optimization.Source: Strategic Recommendations (after human… · p. 88, Overview · p. 89
• HepG2 specificity margin rose by +0.4264 to 0.9367 ± 1.2182, while motif enrichment fell by −0.0212 to 0.9322 ± 0.0682 and DNA stability penalty worsened by +0.0061 to 0.0449 ± 0.0407.The trade-off passage attributes the specificity gain to erosion in motif enrichment and stability; increased margin variance indicates continued exploration.Source: Overview · pp. 88–89
• 50 diverse HepG2-specific enhancers were selected after filtering for DNA stability and ranking candidates by specificity, liver motif enrichment, and HepG2 expression.The selected set means were 1.70 for specificity, 46.83 for liver TF motif enrichment, and 2.73 for HepG2 expression, each above its population mean.Source: Overview · p. 95

S7.4 Inorganic Materials Design

• Iteration 1 raised mean bulk modulus by 179.7 GPa to 270.3 and mean shear modulus by 113.7 GPa to 152.6, but produced no superhard materials.Highly stable oxides tended to be mechanically soft, whereas carbides, borides, and nitrides had higher stiffness and less negative formation energies.Source: Overview · pp. 95–96
• +2.55 GPa brought mean hardness to 18.62 GPa in iteration 2, while the formation-energy penalty improved by 0.364 and the Pugh penalty worsened slightly.Boride-carbides and carbonitrides dominated the upper hardness tail; oxides underperformed.Source: Overview · pp. 96–97
• 32.16 GPa hardness was achieved by Ta3V3B3C4, with good stability and a near-target Pugh ratio.Source: Overview · p. 97
• Analyzer feedback proposed direct hardness optimization, a Pugh-ratio penalty, stronger penalties for unstable compositions, and additional quality objectives such as compactness and toughness.The hardness objective uses the Chen–Niu model; the Pugh penalty targets a ratio centered at 1.6, and the formation-energy penalty weighs positive values more heavily.Source: Overview · pp. 97–99
• For inorganic materials, the Chen hardness surrogate computes predicted Vickers hardness from bulk and shear moduli, using VRH averages when an elastic tensor is available.The specified formula is Hv = max(0, 2∗(k^2)∗G^0.585 − 3), where k = G/B; typical superhard materials have Hv > 40 GPa.Source: Overview · p. 100

S7.5.1 Example of process candidate analysis

• Purity correlated with feed butanol fraction (r=0.79): high-butanol feeds (>0.67) averaged 0.97 purity, versus 0.69 for low-butanol feeds (<0.33).Only 32% of candidates had the ideal two product streams; stream counts ranged from 1–8.Source: S7.5.1 Example of process candidate… · p. 101
• Four candidates (8%) returned a single product essentially equal to the feed, scoring recovery 1.0 despite purity of only 0.46–0.74.The analysis identified these as complete separation failures that expose a loophole in the objective set.Source: S7.5.1 Example of process candidate… · p. 101, Overview · p. 101
• Across 50 feeds, purity averaged 0.7977 ± 0.1368 and recovery 0.6695 ± 0.1718; best purity was 0.9994 and best recovery was 1.0000.The corresponding worst values were 0.4600 for purity and 0.3211 for recovery.Source: Overview · p. 101
• Nine candidates (18%) exceeded purity 0.95 and recovery 0.7, while the best candidate reached purity ~0.998 and recovery ~0.977.All 50 candidates converged in simulation, with performance ranging from trivial failures to near-optimal separations.Source: S7.5.1 Example of process candidate… · p. 100, Overview · p. 101
• The analysis recommended retaining purity as the main objective (weight 1.0), recovery at weight 0.4, and adding a unit-operation complexity penalty weighted 0.1–0.2.It advised deferring component-specific recovery minimums and energy proxies until separation-structure issues are mitigated.Source: Overview · p. 102

S7.5.2 Examples of proposed and implemented objectives

• Balanced component recovery scores the minimum recovery across the two feed components, measuring each component’s recovery into its most enriched product stream.This proposed objective is intended to ensure balanced separation performance.Source: S7.5.2 Examples of proposed and… · p. 102

5. Return min(recoveryA, recoveryB)

• The simplicity scorer gives processes with at most four unit operations a score of 1.0, then applies a hinge penalty as unit count rises.Feed sources and product sinks are excluded from the count.Source: §5 · p. 103
• The recycle-intensity scorer compares component A and B flows in recycle streams with fresh-feed flows, capping the ratio at 1.0.Recycle streams are identified from indices or cycles in the flowsheet graph.Source: §5 · p. 103

SELECTION METHODOLOGY

• The weighted composite score ranked 150 generated flowsheets and selected 50, emphasizing purity while balancing recovery and process simplicity.The selection covered all 49 identified feed compositions.Source: EXECUTIVE SUMMARY · p. 104, SELECTION METHODOLOGY · p. 104

2. SELECTION STRATEGY

• The selection covered all 49 feed compositions with 50 candidates, retaining two for composition (0.4, 0.6, 0.0) and one for each other composition.Iteration 1 supplied 36 selected candidates (72%), iteration 2 supplied 5 (10%), and iteration 3 supplied 9 (18%).Source: §2 · p. 104
• The selection favored iteration 1, which heavily weighted purity and recovery and produced the most successful designs for most feed compositions.Source: §2 · p. 104

SELECTED CANDIDATE PERFORMANCE

• The 50 selected candidates had mean purity 0.9561 and mean minimum component recovery 0.8838; nine achieved a perfect simplicity score of 1.0.The minimum unit-operation simplicity score was 0.3333.Source: SELECTED CANDIDATE PERFORMANCE · pp. 104–105

PERFORMANCE IMPROVEMENT

• Selected processes used more units when simplicity scores fell, while achieving better separation performance.Source: PERFORMANCE IMPROVEMENT · p. 105

ITERATION ANALYSIS

• Iteration 1 achieved perfect simplicity but lower purity (0.71), and its candidates were selected only for five feed compositions where simple processes sufficed.Source: ITERATION ANALYSIS · p. 107
• Iteration 3 improved purity to 0.77 and recovery to 0.84 with variable simplicity of 0.57; its candidates included the top eight overall performers.Source: ITERATION ANALYSIS · p. 107

1. UNIT OPERATIONS

• Distillation columns, decanters, and recycle systems were the primary operations in selected processes.Source: §1 · p. 108
• Most selected processes used 3–8 unit operations.Source: §1 · p. 108
• Most cases used no solvent addition, relying on physical separation, while recycle streams were commonly used for efficiency.Source: §1 · p. 108

2. SEPARATION QUALITY

• Many candidates achieved purity above 99.9.Source: §2 · p. 108
• Product streams showed clear component separation.Source: §2 · p. 108
• Both components were successfully recovered in high purity.Source: §2 · p. 108

• Selected candidates have converged simulations, satisfy mass balances, and represent physically feasible process designs.Source: §3 · p. 108
• All 49 unique feed compositions are represented, ranging from 2% to 98% of the first component.Source: §3 · p. 108
• Each feed composition has its optimal process design selected, with multiple process configurations represented.Source: §3 · p. 108
• The selected processes balance simple designs with 1–4 units and complex designs with 5 or more units.Source: §3 · p. 108

• Candidates achieve the primary goal of high-purity separation and are proposed for detailed process simulation and economic evaluation.Source: §1 · p. 109

• The selected candidates cover the full range of feed compositions, with a proven separation strategy for each composition.Source: §2 · p. 109

3. FUTURE OPTIMIZATION

• Iteration 3 candidates show that balancing all objectives is achievable for specific cases.Source: §3 · p. 109
• Future iterations could target high purity and high simplicity simultaneously and consider multi-objective Pareto optimization.Source: §3 · p. 109

CONCLUSION

• 50 selected flowsheet designs separate binary azeotropic mixtures across diverse feed compositions, with mean product purity of 95.6% and mean component recovery of 88.4%.Iteration 1’s focus on purity was most successful overall; Iteration 3 showed multi-objective balance is achievable for specific cases.Source: CONCLUSION · p. 109